忠実であって修正はしない：マルチホップエージェントリレーにおけるメッセージ形式の影響はティアに依存するFaithful, Not Corrective: Message-Format Effects in Multi-Hop Agent Relays Are Tier-Dependent

agent">Multi-agent systems built on large language models increasingly rely on relay architectures, in which a task passes through a sequence of agents that each read, process, and forward information to the next. A new paper posted to arXiv examines how the format of the messages exchanged between these agents shapes downstream behavior, and reports that the effect is not uniform along the chain but depends on where an agent sits, or its tier. The question matters because relay and pipeline patterns are becoming a common way to decompose complex work, and small design choices in how agents talk to one another can quietly determine whether the overall system is reliable.

The central finding is that message-format effects in multi-hop agent relays are tier-dependent. Rather than acting as quality-control checkpoints, the agents studied tend to propagate the content they receive faithfully, passing information along—including apparent errors—without correcting it. The authors summarize this as behavior that is faithful but not corrective. The practical implication is that once a mistake or a particular formatting choice enters the chain, it is likely to persist and travel downstream rather than being filtered out by later agents. Because the influence of format varies by tier, the same formatting decision can carry different weight depending on how far into the relay it occurs.

This framing connects to a broader concern in agent">multi-agent research often described informally as a telephone-game or error-propagation problem. When outputs from one model become the inputs to another, ambiguities, hallucinations, or brittle structure can compound across hops. The paper's contribution appears to be in isolating message format—for example, whether content is passed as free-form natural language or as more structured representations—as a variable and measuring its downstream effect at different positions rather than treating the chain as a single black box. That per-tier view is useful because it suggests reliability interventions may need to be placed at specific points rather than applied uniformly.

The work sits alongside a growing body of research and tooling around what is sometimes called chain-of-agents, a pattern in which multiple specialized agents collaborate sequentially or hierarchically on long or multi-step tasks. Frameworks such as LangGraph, AutoGen, and CrewAI have made it easier to wire together such pipelines, but they also make it easy to introduce many hand-off points where formatting conventions differ. The finding that upstream agents do not reliably correct upstream content is a caution for developers who might assume that adding more agents, or a downstream reviewer agent, will naturally catch and repair mistakes introduced earlier.

For readers less familiar with the underlying mechanics, a few prerequisite ideas help. In these systems, each agent is typically an LLM guided by a prompt that specifies its role and the expected input and output structure. The message format is the contract between agents, and it interacts with how models parse instructions: structured formats can reduce ambiguity but may also encourage an agent to accept the received content at face value, while looser formats leave more room for reinterpretation. The reported tendency toward faithful propagation is consistent with the observation that models often follow the framing of their input closely, though the paper's specific measurements are what ground this as an empirical result rather than an assumption.

The stated significance is for reliability evaluation in agent">multi-agent system design. If the influence of message format is genuinely tier-dependent, then benchmarking a single agent in isolation may not predict how it behaves inside a relay, and evaluation should account for position within the chain. This points toward design practices such as adding explicit verification steps, defining clearer hand-off contracts, or validating content at chosen tiers rather than trusting that errors will self-correct.

As with any single study, the results should be read with appropriate caution. The findings likely depend on the particular models, tasks, and formats tested, and generalization to other architectures is not guaranteed. Still, the paper adds a useful, more granular lens to a practical problem, reframing message format from a minor implementation detail into a factor whose reliability impact varies systematically across an agent relay.